diff --git a/.agents/skills/testing-arc-solver/SKILL.md b/.agents/skills/testing-arc-solver/SKILL.md
new file mode 100644
index 00000000..a35dff6f
--- /dev/null
+++ b/.agents/skills/testing-arc-solver/SKILL.md
@@ -0,0 +1,101 @@
+# Testing: ARC-AGI Solver & Benchmark Pipeline
+
+## Overview
+The ARC-AGI solver is a bounded symbolic solver in `axioms/arc_solver.py` that uses DSL enumeration and pattern recognition (`axioms/pattern_recognition.py`) to solve ARC-AGI tasks. The benchmark runner (`benchmarks/run_arc_benchmark.py`) runs the solver against full datasets and generates Merkle-anchored evidence.
+
+## Prerequisites
+
+### ARC-AGI Data
+The full ARC-AGI dataset (800 tasks) is needed for benchmark testing:
+```bash
+git clone --depth 1 https://github.com/fchollet/ARC-AGI.git /tmp/ARC-AGI
+```
+This provides:
+- 400 training tasks at `/tmp/ARC-AGI/data/training/`
+- 400 evaluation tasks at `/tmp/ARC-AGI/data/evaluation/`
+
+### Python Dependencies
+```bash
+pip install pytest
+```
+Note: `pytest-timeout` is NOT installed in this repo. Do not use `--timeout` flag.
+
+## Test Commands
+
+### ARC Solver Unit Tests (fastest, run first)
+```bash
+python -m pytest tests/test_arc_solver.py -v
+```
+Expected: 5/5 pass in ~2s. Tests demo task solving, prediction hashes, program compilation, DSL enumeration.
+
+### All PR-Relevant Tests
+```bash
+python -m pytest tests/test_arc_solver.py tests/test_pattern_recognition.py tests/test_ai_invariants.py tests/test_conditional_patterns.py tests/test_cross_model.py tests/test_forgiveness_integration.py -v
+```
+Expected: 15/15 pass in ~5s.
+
+### Benchmark Runner (Demo Mode)
+```bash
+python benchmarks/run_arc_benchmark.py --demo-only --evidence-dir /tmp/test_evidence
+```
+Expected: `10/10 solved (100.00%)`. Writes `benchmark_demo.json` and `manifest_demo.sha256` to the evidence dir.
+
+### Benchmark Runner (Full Dataset)
+```bash
+python benchmarks/run_arc_benchmark.py --data-dir /tmp/ARC-AGI/data --timeout 3 --evidence-dir /tmp/test_evidence
+```
+Expected: ~20/400 training (5%), ~2/400 evaluation (0.5%). Takes ~5-10 minutes.
+
+### Verify a Specific Solved Task
+```python
+import json, sys
+sys.path.insert(0, '.')
+from axioms.arc_solver import load_arc_task_from_json, benchmark_arc_task
+
+with open('/tmp/ARC-AGI/data/training/1cf80156.json') as f:
+    data = json.load(f)
+task, expected = load_arc_task_from_json('1cf80156', data)
+solved, proof = benchmark_arc_task(task, expected, max_depth=3)
+print(f'Solved: {solved}, Proof valid: {proof.is_valid()}')
+```
+
+### Verify Evidence Chain Integrity
+```python
+import json, hashlib
+with open('evidence/arc_agi_3/benchmark_demo.json') as f:
+    data = json.load(f)
+payload = json.dumps(
+    [{'task_id': r['task_id'], 'solved': r['solved'], 'prediction_hash': r['prediction_hash']} for r in data['task_results']],
+    sort_keys=True, separators=(',', ':'),
+)
+computed = hashlib.sha256(payload.encode('utf-8')).hexdigest()
+print(f'Match: {computed == data["manifest_hash"]}')
+```
+
+## Full Test Suite
+```bash
+python -m pytest tests/ -v --ignore=tests/test_f_platform_001.py --ignore=tests/test_f_platform_002.py --ignore=tests/test_f_platform_003.py --ignore=tests/test_f_platform_004.py --ignore=tests/test_f_platform_005.py --ignore=tests/test_pr45_uvdtl.py
+```
+
+### Known Pre-Existing Failures (Not Related to ARC Solver)
+These tests fail due to missing build artifacts (`out/` directory) or missing modules. They are NOT caused by ARC solver changes:
+- `test_f_platform_001-005.py` — `ModuleNotFoundError: No module named 'test_falsification'` (collection error, must be `--ignore`d)
+- `test_pr45_uvdtl.py` — `ModuleNotFoundError: No module named 'pr45_uvdtl.build'` (collection error, must be `--ignore`d)
+- `test_food_cart_universe.py` — missing `out/food_cart_dag.json`
+- `test_skate4_multiplayer_universe.py` — missing `out/skate4_mp_dag.json`
+- `test_uncharted_multiplayer_universe.py` — missing `out/uncharted_mp_dag.json`
+- `test_successor_readiness.py` — missing `canonical/hash_manifest.json`
+- `test_pr34_audit.py`, `test_pr50_bar_exam.py`, `test_repository_compliance.py` — various pre-existing issues
+
+## Key Architecture Notes
+- `_select_best_rule()` in `pattern_recognition.py` validates full operation chains before MDL tiebreaker
+- Multi-object decomposition is wired via `_infer_per_object_rule` in `_candidate_rules_for_pair`
+- Evidence files use SHA-256 manifest hashes computed from JSON-serialized task results
+- The benchmark runner uses `signal.SIGALRM` for per-task timeouts (Linux only)
+
+## CI Notes
+- CI workflow (`constitution.yml`) may not include `test_arc_solver.py` — OAuth `workflow` scope might be needed to push changes to this file
+- 27 CI checks run on push; all should pass
+
+## Devin Secrets Needed
+None — this is a pure Python library with no external service dependencies.